Clef

Clef is a 27B open-weight decision model. Like Sarvam Prism, it answers typed questions about an input on the Decisions API: you send a state and a map of questions, and it returns a probability for every option of every question in one pass. It generates no text, so there is nothing to parse.

Unlike Sarvam Prism, Clef also reads images, and it takes up to 65,536 tokens per request.

At a glance

Quickstart

from sarvamai import SarvamAI
client = SarvamAI(api_subscription_key="YOUR_SARVAM_API_KEY")
response = client.decisions.create(
    model="clef",
    state="My Stripe integration stopped working and all checkout payments are failing.",
    questions={
        "urgent": {
            "type": "noul",
            "instructions": "Is this urgent and revenue-impacting?",
        },
        "department": {
            "type": "choice",
            "instructions": "Which team should handle this?",
            "criteria": {
                "billing": "Payments or invoices",
                "technical": "Bugs or outages",
                "sales": "New business enquiries",
            },
        },
        "severity": {
            "type": "score",
            "instructions": "Rate the severity.",
            "criteria": ["trivial", "minor", "moderate", "major", "critical"],
        },
    },
)
answers = response.answers
if answers["urgent"].noul > 0.5:
    print("Escalate to", answers["department"].choice)

Response:

{
  "model": "clef",
  "answers": {
    "urgent": {"type": "noul", "noul": 0.9926},
    "department": {
      "type": "choice",
      "choice": "billing",
      "confidence": 0.3099,
      "probabilities": {"billing": 0.5400, "technical": 0.4476, "sales": 0.0124}
    },
    "severity": {
      "type": "score",
      "score": 3.8746,
      "confidence": 0.8955,
      "legend": {"0": "trivial", "1": "minor", "2": "moderate", "3": "major", "4": "critical"},
      "probabilities": {"0": 0.0071, "1": 0.0068, "2": 0.0089, "3": 0.0587, "4": 0.9184}
    }
  },
  "usage": {"input_tokens": 357, "output_tokens": 0}
}

Probabilities are rounded here; the API returns full floats. For what each field means, see Read the answers.

Questions about images

Send images in images, as base64 data URIs. The questions are answered about the state and the images together. Remote http(s) image URLs return 400.

The sarvamai SDK’s decisions.create doesn’t have an images argument yet, so pass the images as an extra body field:

import base64
with open("receipt.jpg", "rb") as f:
    image = "data:image/jpeg;base64," + base64.b64encode(f.read()).decode()
response = client.decisions.create(
    model="clef",
    state="Review the attached receipt.",
    questions={"legible": {"type": "noul", "instructions": "Is the receipt total legible?"}},
    request_options={"additional_body_parameters": {"images": [image]}},
)
print(response.answers["legible"].noul)

In the JavaScript SDK, put images directly in the request object: client.decisions.create({ model: "clef", state, questions, images: [image] }). Its requestOptions can’t add body fields.

Over HTTP, send images in the request body:

{
  "model": "clef",
  "state": "Review the attached receipt.",
  "images": ["data:image/jpeg;base64,/9j/4AAQ..."],
  "questions": {
    "legible": {"type": "noul", "instructions": "Is the receipt total legible?"}
  }
}

Question rules

The question types are the same as on Sarvam Prism (Question types), with these rules on Clef:

• Always send instructions (or noul criteria). The question ID is not used as the question.
• Any key in a question other than type, instructions, and criteria returns 422, including misspellings such as instruction.
• temperature returns 422. max_tokens and stream are accepted and ignored: Clef always answers in one JSON response.
• chat_template_kwargs must be empty or absent; a non-empty value returns 400.
• videos and other unknown top-level fields are ignored.

Limits

• 65,536 tokens per request, counting the state, every question and its options, and images. The state is counted once, however many questions you ask.
• The state is never truncated. A request over the limit fails with 400 invalid_request_error, and the message gives the count, such as “the prompt needs 70141 tokens (70001 for the state), but holds at most 65536”. Shorten the state, or split a long document across requests.
• Ask several questions about the same state in one request: they are answered together, in one pass.

Differences from Sarvam Prism

Errors

Every error carries message, code, and request_id; validation errors also carry param.